Understanding the decision-making process of deep learning networks is a key challenge that has rarely been investigated for synthetic aperture radar (SAR) images. In this article, a set of new analytical tools is proposed and applied to a convolutional neural network (CNN) handling automatic target recognition on two SAR datasets containing military targets. First, an analysis of the respective influence of target, shadow, and background areas on classification performance is carried out. The shadow appears to be the least used portion of the image affecting the decision process, compared to the target and clutter, respectively. Second, the location of the most influential features is determined with classification maps obtained by systematically hiding specific target parts and registering the associated classification rate relative to the images to be classified. The location of the image areas without which classification fails is target type and orientation specific. Nonetheless, a strong contribution of specific parts of the target, such as the target top and the areas facing the radar, is noticed. Finally, results show that features are increasingly activated along the CNN depth according to the target type and its orientation, even though target orientation is absent from the loss function.
Explainability of Deep SAR ATR Through Feature Analysis
IEEE Transactions on Aerospace and Electronic Systems ; 57 , 1 ; 659-673
2021-02-01
4349159 byte
Article (Journal)
Electronic Resource
English
Driving through the Concept Gridlock: Unraveling Explainability Bottlenecks in Automated Driving
ArXiv | 2023
|